Skip to contentSkip to conversation

crypto: hisilicon/zip - Optimize performance by replacing rw_lock with spinlock

PATCH In review Kuan-Wei Chiu 2 participants 2 messages Aug 23, 2024 – Aug 30, 2024

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 crypto: hisilicon/zip - Optimize performance by replacing rw_lock with spinlock

The req_lock is currently implemented as a rw_lock, but there are no instances where read_lock() is called. This means that the lock is effectively only used by writers, making it functionally equivalent to a simple spinlock.

As stated in Documentation/locking/spinlocks.rst:
"Reader-writer locks require more atomic memory operations than simple spinlocks. Unless the reader critical section is long, you are better off just using spinlocks."

Since the rw_lock in this case incurs additional atomic memory operations without any benefit from reader-writer locking, it is more efficient to replace it with a spinlock. This patch implements that replacement to optimize the driver's performance.

---
Note: Build test only.

 drivers/crypto/hisilicon/zip/zip_crypto.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
drivers/crypto/hisilicon/zip/zip_crypto.c +7 -7
index 94e2d66b04b6..92d3bd0dfe1b 100644
@@ -54,7 +54,7 @@ struct hisi_zip_req {
 struct hisi_zip_req_q {
 	struct hisi_zip_req *q;
 	unsigned long *req_bitmap;
-	rwlock_t req_lock;
+	spinlock_t req_lock;
 	u16 size;
 };
 
@@ -116,17 +116,17 @@ static struct hisi_zip_req *hisi_zip_create_req(struct hisi_zip_qp_ctx *qp_ctx,
 	struct hisi_zip_req *req_cache;
 	int req_id;
 
-	write_lock(&req_q->req_lock);
+	spin_lock(&req_q->req_lock);
 
 	req_id = find_first_zero_bit(req_q->req_bitmap, req_q->size);
 	if (req_id >= req_q->size) {
-		write_unlock(&req_q->req_lock);
+		spin_unlock(&req_q->req_lock);
 		dev_dbg(&qp_ctx->qp->qm->pdev->dev, "req cache is full!\n");
 		return ERR_PTR(-EAGAIN);
 	}
 	set_bit(req_id, req_q->req_bitmap);
 
-	write_unlock(&req_q->req_lock);
+	spin_unlock(&req_q->req_lock);
 
 	req_cache = q + req_id;
 	req_cache->req_id = req_id;
@@ -140,9 +140,9 @@ static void hisi_zip_remove_req(struct hisi_zip_qp_ctx *qp_ctx,
 {
 	struct hisi_zip_req_q *req_q = &qp_ctx->req_q;
 
-	write_lock(&req_q->req_lock);
+	spin_lock(&req_q->req_lock);
 	clear_bit(req->req_id, req_q->req_bitmap);
-	write_unlock(&req_q->req_lock);
+	spin_unlock(&req_q->req_lock);
 }
 
 static void hisi_zip_fill_addr(struct hisi_zip_sqe *sqe, struct hisi_zip_req *req)
@@ -456,7 +456,7 @@ static int hisi_zip_create_req_q(struct hisi_zip_ctx *ctx)
 
 			goto err_free_comp_q;
 		}
-		rwlock_init(&req_q->req_lock);
+		spin_lock_init(&req_q->req_lock);
 
 		req_q->q = kcalloc(req_q->size, sizeof(struct hisi_zip_req),
 				   GFP_KERNEL);
Signature · 1 line
2.34.1
maintainer
#1 lore raw
Quoted 15 lines · Kuan-Wei Chiu

The req_lock is currently implemented as a rw_lock, but there are no instances where read_lock() is called. This means that the lock is effectively only used by writers, making it functionally equivalent to a simple spinlock.

As stated in Documentation/locking/spinlocks.rst:
"Reader-writer locks require more atomic memory operations than simple spinlocks. Unless the reader critical section is long, you are better off just using spinlocks."

Since the rw_lock in this case incurs additional atomic memory operations without any benefit from reader-writer locking, it is more efficient to replace it with a spinlock. This patch implements that replacement to optimize the driver's performance.

Quoted 6 lines

---
Note: Build test only.

 drivers/crypto/hisilicon/zip/zip_crypto.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

Patch applied. Thanks.

Signature · 3 lines
Email: Herbert Xu <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt